Two-dimensional reduction PCA: a novel approach for feature extraction, representation, and recognition

Abstract

We develop a novel image feature extraction and recognition method two-dimensional reduction principal component analysis (2D-RPCA)). A two dimension image matrix contains redundancy information between columns and between rows. Conventional PCA removes redundancy by transforming the 2D image matrices into a vector where dimension reduction is done in one direction (column wise). Unlike 2DPCA, 2D-RPCA eliminates redundancies between image rows and compresses the data in rows, and finally eliminates redundancies between image columns and compress the data in columns. Therefore, 2D-RPCA has two image compression stages: firstly, it eliminates the redundancies between image rows and compresses the information optimally within a few rows. Finally, it eliminates the redundancies between image columns and compresses the information within a few columns. This sequence is selected in such a way that the recognition accuracy is optimized. As a result it has a better representation as the information is more compact in a smaller area. The classification time is reduced significantly (smaller feature matrix). Furthermore, the computational complexity of the proposed algorithm is reduced. The result is that 2D-RPCA classifies image faster, less memory storage and yields higher recognition accuracy. The ORL database is used as a benchmark. The new algorithm achieves a recognition rate of 95.0% using 9 × 5 feature matrix compared to the recognition rate of 93.0% with a 112 × 7 feature matrix for the 2DPCA method and 90.5% for PCA (Eigenfaces) using 175 principal components.